Skip to content

Foundation Model-driven Key Anatomy Frame Selection for Blind-sweep Ultrasound Fetal Birth Weight Estimation

🕒 Published (v1): 2026-07-01 10:27 UTC · Source: Arxiv · Venue: MICCAI 2026 · link

Why this paper was selected

MICCAI 2026; foundation model for ultrasound frame selection; low-resource medical VLM

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper presents a foundation model-driven framework for estimating fetal birth weight (FBW) from blind-sweep ultrasound videos — freehand scans acquired without requiring standard anatomical plane identification — enabling operator-independent prenatal assessment. The method uses FetalCLIP, a vision-language foundation model, to select anatomically informative frames from noisy, unconstrained video sequences, followed by redundancy-aware compression and prior-guided regression. On 839 patients it achieves MAE of 161.3 g, surpassing clinical Hadlock-4 estimation.

Problem

Existing AI-based fetal weight estimation methods require standardized ultrasound planes and expert sonographers to acquire high-quality input data, limiting deployment in low-resource settings. No prior work applied the operator-independent blind-sweep paradigm — continuous freehand abdominal scanning without real-time plane identification — to fetal birth weight regression, leaving a critical clinical gap for settings with few experienced sonographers.

Method

The framework processes six blind-sweep loop videos per exam through three sequential modules:

  1. Anatomy-Guided Frame Selection (AFS): Uses FetalCLIP (a CLIP-style image-text model pretrained on fetal ultrasound data) to embed all video frames \(\{v_1,\ldots,v_N\}\) and fixed anatomical text prompts \(\{t_1,\ldots,t_M\}\) (head, abdomen, femur) into a shared embedding space. Per-frame anatomical confidence is \(c = \max_j \cos(X, P)_{:,j} \in \mathbb{R}^N\). The top-\(K=100\) frames are retained as a coarse subset \(X_c\).

  2. Redundancy-Aware Feature Compression (RAFC): A lightweight adapter transforms \(X_c\) into \(\tilde{X}\); a Transformer layer produces frame-level representations \(H\). Task-specific importance scores are \(s_\text{imp} = \sigma(W_\text{imp}H)\). Local diversity scores penalize redundancy within a sliding window of \(w=10\): \(d_i = 1 - \max_{j\in W_i}\cos(h_i, h_j)\). A gating MLP fuses \(s_\text{imp}\), diversity \(s_\text{div}\), and AFS confidence \(c\) into selector scores; the top-\(L=50\) frames per loop are retained.

  3. Prior-Guided Attention Regressor (PGAR): Frame selector scores define a softmax prior \(p_i = \text{softmax}(a_i/\tau)\) used to reweight features \(\tilde{H}_f = p \times H_f\) before self-attention, global average pooling, and a regression head producing \(\hat{y}\).

Features from all six loops (\(6L = 300\) frame features) are concatenated for a single FBW prediction. FetalCLIP encoders are frozen; only the adapter and task-specific layers are trained.

Key Contributions

  • First application of the blind-sweep ultrasound paradigm to fetal birth weight estimation, decoupling accuracy from operator expertise.
  • AFS module using FetalCLIP vision-language alignment for anatomy-guided, coarse keyframe selection from unconstrained video.
  • RAFC module combining task-importance, local diversity, and anatomical confidence for fine-grained redundancy suppression.
  • PGAR that injects the learned frame-relevance prior into attention-based aggregation before regression.
  • Prospectively collected dataset of 870 exams from 839 patients with post-delivery ground-truth weights.

Results

  • MAE 161.34 g, MAPE 4.89%, PC10 90.23%, PC15 100.0% (test set, \(n=174\) exams).
  • Outperforms clinical Hadlock-4 (MAE 183.57 g, PC10 82.18%, PC15 97.13%) — the only competing method surpassing the clinical standard.
  • Outperforms INTERGROWTH-21st (MAE 216.63 g), USFM (MAE 247.26 g), SelectGA (MAE 287.03 g), EchoNet (MAE 295.72 g), ViFi-CLIP (MAE 284.94 g), ResNet50+WAA (MAE 309.50 g).
  • Only method meeting clinical acceptance thresholds (PC10 ≥ 80%, PC15 ≥ 95%).
  • Ablation: AFS alone reduces MAE by 18.8% and raises PC10 by 10.5% over baseline; RAFC alone by 27.5% MAE and 11.2% PC10; combined yields further 8.7%/6.6% gains.

Limitations

  • Single-center dataset; authors explicitly note multi-center validation is future work.
  • Blind-sweep protocol is standardized (fixed number/trajectory of sweeps), so operator-independence is partial — protocol adherence is still required.
  • Frame selection hyperparameters (\(K\), \(L\), \(w\)) were tuned on a single validation split; generalizability untested.
  • Dataset covers a narrow birth-weight range (1410–4360 g, mean 3282 g); performance on growth-restricted or macrosomic extremes is unknown.
  • Only six-loop exam configuration evaluated; robustness to varying sweep counts not assessed.

Relevance to Vision-Language Models

FetalCLIP — a domain-specific VLM pretrained on fetal ultrasound image-text pairs — is the linchpin of the AFS module: its shared embedding space enables zero-shot anatomical confidence scoring without any frame-level annotation, demonstrating that VLMs can replace explicit plane-detection modules in clinical video pipelines. This is a concrete downstream deployment case showing how a domain-adapted CLIP model transfers to a fine-grained, weakly supervised regression task with frozen encoders. The work also highlights a general design pattern for medical VLMs: use cross-modal similarity as a structural prior to filter uninformative inputs before task-specific heads, which may generalize to other diagnostic imaging modalities.